Every team building their own agents today asks the same question before they can get to production: what is this going to cost us?
Token pricing makes the answer somewhat nuanced. Depending on whether your agents run on OpenAI, Anthropic, Google, or another provider, a single workflow may burn a few hundred tokens on a quick classification, or several hundred thousand on a lengthy document analysis with knowledge retrieval and tool calls. Multiply that across a few teams and a few dozen workflows, and your monthly bill moves by an order of magnitude.
Here's how to think about where your tokens go, and what to do about it.
Why token bills grow faster than teams expect
The first issue is the cultural shift. High token consumption has become a proxy for sophistication: a sign that your agents are "really working," that you're doing serious AI via "tokenmaxxing." Teams benchmark against each other on usage, leadership thinks that burn rate is evidence of adoption, and the biggest bill starts to feel like special status. But it isn't.
Tokens are not an outcome. A workflow that chews through half a million tokens isn't more productive than one that does the same job in fifty thousand, it's just more expensive. Merely spending more on tokens doesn't mean that AI is closing support tickets, processing documents, or relieving administrative burden.
Over the past few years, I've noticed that there also tend to be three technical issues driving token overspend.
The first is reaching for the fanciest model at every step. Frontier models feel exciting and cutting-edge, so teams route everything through GPT-5-class or Claude Sonnet-level reasoning. In practice, maybe one or two steps in a workflow need that. Classification, extraction, formatting, and routing run fine on models that cost 20 to 50 times less per token.
The second is no visibility into per-step or per-agent token consumption. If you can't see at a granular level which agent is spending the most tokens, and how it responds to different kinds of inputs, you can't fix it.
The third is that context keeps growing on its own. RAG pipelines and long chat histories keep expanding the context window. A chatbot that remembers every prior message costs roughly ten times more per message by month six than it did when first published.
The playbook for managing enterprise AI tokens
Having worked with hundreds of CIOs to transform their enterprises with AI, I’ve found there are a few key steps to implement when evaluating token usage.
Choose the right-size model for each task. Use small, fast models (Haiku-tier, GPT-5-mini-tier, GLM 5.2-tier) for routing, classification, and simple transforms. Save the frontier models for the one or two steps that need best-in-class reasoning.
Keep context tight. Chunk documents instead of point-blank dumping them in and use structured outputs so responses stay short.
Monitor tokens at a granular level. Orchestrate every workflow so you know where tokens go and the monthly bill doesn't shock you. Better yet, create a single-pane-of-glass dashboard for agent analytics that break down how many tokens were used at each step.
Set daily limits at different levels: org, user, and project. Guardrails like this can help stop a runaway agent (or a curious intern) from churning through a month's budget in an afternoon.
How we think about tokens at StackAI
Even if you follow all of the above steps, though, there’s still one caveat: most platforms that sit between you and the model providers also charge by tokens, which means every optimization you make lowers their revenue. They have no reason to help you use fewer tokens, and plenty of reason to leave the default set to the expensive model.
That’s why we made the deliberate choice to bill per run at StackAI. A run is defined as one complete end-to-end execution of a workflow, regardless of the tokens consumed. If your product handles roughly 50,000 support conversations a month, you know your StackAI cost before the month starts. A longer support question doesn't move the number, and nobody on finance is going to send off angry emails at unexpected overages.
But the number on the invoice was never the real test. Instead, we ask: are you getting the most value out of your tokens? A billion tokens spent on inefficient, suboptimal workflows is “Tokenmaxxing.” The same tokens routed through workflows dedicated to specific tasks is what we call “10x Tokens”: 10x the value for the same spend, which is what we achieve at StackAI. Applying every token to the right step of an agentic process is how you stop wasting them.
You don't have to go through this alone. Every StackAI customer gets a solutions team (an AI strategist and a forward-deployed engineer) that's highly experienced in helping enterprises adopt AI safely and effectively. They'll audit your use cases, debug your workflows, and find the three to five nodes that could be further optimized. They'll help you navigate analytics, Gantt charts, per-agent token consumption, and more. Our customers note significant savings with StackAI after switching over from homegrown agents.
We’re proud to stand for real AI adoption amidst the rising tide of tokens for tokens’ sake—choosing to bill per run, include the tokens, and get a solutions team on your side whose job is to help you get the best value with AI.
Want to learn more about how StackAI can support your enterprise AI journey? Get a demo here.
Bernard Aceituno
Co-Founder at StackAI
